Taming the Long Tail: Rebalancing Adversarial Training via Adaptive Perturbation

Researchers propose RobustLT, a plug-and-play adversarial training framework for long-tailed datasets.

The paper shows that class imbalance and unstable adversarial distributions are major obstacles to robust long-tailed learning.

RobustLT adaptively adjusts perturbations to improve both adversarial robustness and balance across classes.

The method targets a common real-world setting where imbalanced data and adversarial vulnerability appear together.